Sensitivity Analysis Method Of Bayesian Networks And Its Application On Fault Diagnosis Of Complex System

In the past few years, the research about uncertainty knowledge system ofAI(Artificial Intelligence) make great progress, Bayesian Networks, as a tool toexpress uncertainty knowledge has also made great progress, and sensitivityanalysis of Bayesian Networks has become a hot spot in the practical application.The sensitivity analysis of Bayesian networks is to research on the relation betweenparameters or evidence and a conditional probability of object node. Nowadays,sensitivity analysis has wide applications in medicine, civil engineering, computer,fault diagnosis and etc. But there is no sensitivity algorithm about DynamicBayesian Networks. So, we propose a sensitivity analysis algorithm about DynamicBayesian Networks based on FF (Factored Frontier), and the sensitivity analysis isused for dealing with fault diagnosis in complex systems. The main contents of thisdissertation are as follows:(1) This dissertation makes a survey about the research on sensitivity analysisof Bayesian Networks and fault diagnosis, including the background, the currentresearch state.(2) The sensitivity analysis methods of HMMs model can not be used forgeneral DBNs(Dynamic Bayesian Networks), and there is high complexity. A newalgorithm(SA_FF algorithm) is presented, which can effectively deal withsensitivity analysis problems of DBNs. Based on FF algorithm, SA_FF algorithmcan calculate the sensitivity function of regular DBNs. The relation between givenparameters and object node distribution is established by dynamic reason forfrontier nodes. The SA_FF algorithm significantly improves computationalefficiency by the marginalization of local frontier rather than updating the jointprobability distribution of model. On the other hand, the SA_FF algorithm can beused for multiple parameters sensitivity analysis of DBNs; although some error isinduced, the error is bounded by demonstration. And the effectiveness of SA_FFalgorithm is illustrated by some examples.(3) The DFC algorithm only consider their own state of probability, increasedthe time for looking abnormal nodes, leading to the tine complexity of thealgorithm is too high. Apply the sensitivity analysis with fault diagnosis andpropose fault diagnosis algorithm: SA_FD algorithm. The states of the nodecombine with the state parameter for influence the child node states by means of calculating the sensitivity function, therefore, we can looking for the best importantnode efficiently, so shorten the time to looking for abnormal node, and we will findthe fault node by the fastest efficiency so as to improve the efficiency of the faultdiagnosis.